llama.cpp b10322 coalesces memory loads for SSM convolution on Intel SYCL, delivering up to 1.87x speedup on Arc Pro B70 for state-space-model operations.
The release ships one change: PR #26612, which optimizes the ssm_conv window load pattern. Benchmarks on an Intel Arc Pro B70, interleaved A/B against master across 6 reps:
- SSM_CONV, 512 tokens: 97.68 → 52.95 µs/run (1.85x faster)
- SSM_CONV, 934 tokens: 516.16 → 276.13 µs/run (1.87x faster)
- SSM_CONV, 1 token: 2.73 → 2.71 µs/run (flat)
On Qwen 3.5 27B Q4_K Medium — where 48 of 64 blocks run ssm_conv — end-to-end prompt processing gains are modest:
pp2048 -b 2048 -ub 2048: +2.2% (1045 → 1070 t/s)pp2048 -b 2048 -ub 512: +1.8% (772 → 786 t/s)tg128: flat at ~23.9 t/s
The gap between raw kernel speedup (1.85x) and end-to-end gains (+2.2%) tells you ssm_conv is one operation among many in the pipeline. The convolution kernel got faster, but it’s not the bottleneck for overall throughput.
NOTE
This optimization targets the SYCL backend (Intel GPUs) only. CUDA, Metal, and Vulkan backends are unaffected. Token generation is flat across all configs — gains are specific to prompt processing where ssm_conv runs in batch.
Available now in llama.cpp b10322. Prebuilt binaries cover macOS Apple Silicon, Ubuntu x64/arm64 (CPU, Vulkan, ROCm 7.2, OpenVINO), and s390x. The KleidiAI-enabled macOS build is disabled in this release.